#Architecture & Core Concepts

This chapter covers LangChain 1.0 layering, the agent loop, and when to use LangGraph.

#Layered stack

┌─────────────────────────────────────────┐
│  Your app: API, UI, jobs                │
├─────────────────────────────────────────┤
│  LangChain: create_agent, middleware    │
├─────────────────────────────────────────┤
│  LangGraph: StateGraph, checkpointer    │
├─────────────────────────────────────────┤
│  langchain-core: messages, tools, Runnable│
├─────────────────────────────────────────┤
│  Integrations: openai, anthropic, …     │
└─────────────────────────────────────────┘

• langchain-core: vendor-neutral abstractions
• LangGraph: stateful, loop-capable runtime with persistence
• LangChain 1.0: standard agent loop + middleware on top of LangGraph

#Agent = Model + Harness

The harness delivers the right context at the right time.

from langchain.agents import create_agent

agent = create_agent(
    model="openai:gpt-4.1",
    tools=[...],
    system_prompt="...",
    middleware=[...],
    checkpointer=...,
)

#Loop on LangGraph

create_agent builds a LangGraph graph:

• model node: LLM call, may emit tool_calls
• tools node: run tools, append results to messages

Middleware wraps model entry/exit like web middleware.

#LangChain vs LangGraph

Guidance: start with LangChain; drop to LangGraph for custom orchestration; embed create_agent as a subgraph when needed.

#Workflows vs agents

See LangGraph Workflows for patterns.

#Runnable & LCEL

Pre-1.0 LCEL composed Runnables with |. Runnables remain the low-level protocol; agents are the main 1.0 path via create_agent. Integration classes like ChatOpenAI are still Runnables usable in graph nodes.

#Versions

• Align langchain 1.x with langgraph 1.x
• Avoid legacy AgentExecutor / initialize_agent in new projects
• Docs: docs.langchain.com

#Next steps

• Models & Messages
• Tools & Agents
• Middleware